Two Neural Features Related to Information Encoding and Behavior
Post by Stephanie Williams
What's the science?
To understand how neurons encode information related to the external world, such as information from the environment, we need to understand which statistical features of individual neurons contain that information. In the past, research groups have suggested the mean firing rate of individual neurons or groups of neurons may represent information about stimuli. Other research has looked at the amount of noise that populations of neurons share (correlated noise), however, the specific statistical properties of neurons related to the information encoding are not well understood. This week in The Journal of Neuroscience, Noguiera and colleagues identify neural features that explain most of the variance in information encoding and behavior. The authors specifically address two questions 1) what features are related to information encoding 2) do those features affect behavioral performance?
How did they do it?
This study involved both experimental data collection and theoretical modelling. For the experimental arm of the study, four monkeys were trained to perform three different tasks. Two of the three tasks were direction discrimination tasks (one coarse discrimination and one fine discrimination), and the third was a spatial attention task, in which two of the monkeys had to detect a change in the orientation of some lines displayed in a circle (i.e. a Gabor patch). The authors recorded neural activity while the monkeys performed a given task from two brain regions: the middle temporal area (MT) and area 8a in the lateral prefrontal cortex. Performance on the tasks was quantified as the number of correct reports of motion for the direction tasks, and as the mean reaction time for the attention task. To test which neural features were related to 1) information encoding and 2) behavioral performance, the authors isolated features of interest by iterating through each extracted feature, and changing the values of one set of features while holding all of the other features constant. By using a statistical technique, called bootstrapping, they could select the bootstrap iterations that produced feature values that were in a narrow range around the median for that particular feature. They used these bootstrapped samples to generate the “fluctuations” in the feature that they were changing during their iterations (they call this method “conditioned bootstrapping”). The authors then trained a binary classifier to predict which task the monkeys were performing and the specific behavior performed in the task (i.e. which Gabor patches were attended to or left vs. right motion stimulus).
For the theoretical arm of the study, the authors defined decoding performance, mathematically simulated experimental data, and tested decoding performance on the simulated data. To define decoding performance, the authors first derived a mathematical expression that described the theoretical optimal performance of a linear classifier. They defined two terms representing the two categories of features of interest 1) the population signal feature, which is a measure of how the overall modulation of activity of the measured neuronal population changes as a function of the stimulus condition, and 2) the projected precision feature which is related to the trial-by-trial variability. To simulate population activity, the authors built a neural population activity model with a large ensemble than was recorded experimentally (N = 1000 model neurons), with each neuron’s activity modeled as a function of stimuli (2 stimuli) of different strengths (3 different strengths). They also incorporated a mathematical term that represented noise correlations (corresponding changes in trial-to-trial variability) between neurons. To model behavioral performance, the authors used an optimal linear classifier to make predictions from simulated neural activity. They then compared the performance of their theoretical decoder with the performance of the decoder trained on their experimental data.
What did they find?
The authors found that two of the features they examined were important features that affected how much information was encoded, and were the strongest predictors of behavioral performance. They found that changing both the 1) population signal feature and the 2) projected precision feature (one at a time, while holding all other features constant), significantly affected the amount of encoded information and also predicted changes in behavioral performance.
The first feature related to population tuning was specifically a length metric that joined the mean population responses across different experimental conditions. The second feature, related to the amount of trial-by-trial variability was calculated as the inverse of the population covariability projected onto the direction of the population signal. Importantly, the authors did not find that other features (such as global activity and mean pairwise correlations), which had previously been suggested by other research, were related to the amount of information encoding when they controlled for the two features they identified. However, it is worth noting that changing the global activity and correlation features did change the amount of information encoded in population activity when the two features that authors identified were not controlled for.
What's the impact?
The authors show for the first time that two features, population signal and projected precision, modulate the amount of information encoded by finite neuronal populations and predict changes in behavior. They also show that two other features did not modulate the amount of encoded information or behavioral performance. These findings shed light on the specific properties of neurons involved in encoding information.
Nogueira et al. The effects of population tuning and trial-by-trial availability on information encoding and behavior. J. Neurosci (2019). Access the original scientific publication here.